Grok Build setup (Grok 4.5)

How to run Grok Build CLI/TUI against the gateway with grok-4.5.

Recommended settings

Item	Value
Base URL	`https://sub2api.lingshichat.top/v1`
model	`grok-4.5`
api_backend	`responses`
supports_backend_search	`true`
web_search model	Also `grok-4.5` (separate model section)
context_window (practical)	`200000` (~200k)
context_window (hard max)	Model accepts up to `500000` (~500k); do not run there day-to-day
Data training / improve the model	Off (required)

Grok 4.5 context: 500k max, but quality drops past ~200k.
The model can accept about 500k tokens, but reasoning quality degrades clearly after ~200k (worse answers, missed edits, flaky tool use).
For coding agents, set client context_window to 200000 so auto-compaction runs before that cliff. Do not keep multi-hour sessions just to “use the full 500k”.

Turn off Grok Build data training / “improve the model” data sharing.
Using a relay does not skip client privacy settings. Before configuring models, run /privacy and confirm training / data sharing is off. See Disable data training.

If chat works but web_search returns 402 Payment Required / spending-limit, search is often using a separate model without backend search. Point both chat and web_search models at gateway grok-4.5 and set supports_backend_search = true.

0. Disable data training (required)

Using the relay does not mean you can skip client privacy settings.

Start Grok Build (TUI).
Run:

/privacy

Review privacy / data-retention status.
Turn off options related to training / improve-the-model / data sharing for training.
Run /privacy again to confirm.

You can also open settings:

/settings

Find Privacy / Data controls (or similar) and disable “Improve the model” style toggles.

Optional: disable client telemetry

In ~/.grok/config.toml:

[features]
telemetry = false

This is separate from data-training toggles; both are worth reviewing.

/privacy controls privacy / retention / training participation; [features].telemetry controls anonymous usage telemetry.
If you still use an xAI / X account, also disable Grok training options under X → Settings → Privacy and safety → Grok.

1. Prepare an API key

Create a key in the console
Prefer an environment variable:

# Windows PowerShell
$env:ARONA_API_KEY = "sk-xxxxxxxx"
[System.Environment]::SetEnvironmentVariable("ARONA_API_KEY", "sk-xxxxxxxx", "User")

# macOS / Linux
export ARONA_API_KEY="sk-xxxxxxxx"

2. Edit `~/.grok/config.toml`

Windows:

C:\Users\<you>\.grok\config.toml

macOS / Linux:

~/.grok/config.toml

Recommended full config

[models]
default = "arona-grok45"
web_search = "arona-web-search"
# default_reasoning_effort = "high"

[model.arona-grok45]
model = "grok-4.5"
name = "Arona Grok 4.5"
description = "via sub2api.lingshichat.top"
base_url = "https://sub2api.lingshichat.top/v1"
env_key = "ARONA_API_KEY"
api_backend = "responses"
# Practical window ~200k. Hard max ~500k, but quality drops past ~200k
context_window = 200000
supports_backend_search = true

[model.arona-web-search]
model = "grok-4.5"
name = "Arona Grok 4.5 Web Search"
base_url = "https://sub2api.lingshichat.top/v1"
env_key = "ARONA_API_KEY"
api_backend = "responses"
context_window = 200000
supports_backend_search = true

Option B: override built-in `grok-4.5`

[models]
default = "grok-4.5"
web_search = "grok-4.5"

[model.grok-4.5]
base_url = "https://sub2api.lingshichat.top/v1"
env_key = "ARONA_API_KEY"
api_backend = "responses"
context_window = 200000
supports_backend_search = true

When overriding a built-in model, only set fields you change. Still set supports_backend_search = true for web_search.

Option C: shared endpoint base

[endpoints]
models_base_url = "https://sub2api.lingshichat.top/v1"

[models]
default = "grok-4.5"
web_search = "grok-4.5"

[model.grok-4.5]
env_key = "ARONA_API_KEY"
api_backend = "responses"
supports_backend_search = true
context_window = 200000

With models_base_url, Grok can list models via API key auth and usually does not need grok login.

3. Launch and switch models

grok models
grok -m arona-grok45
# in session:
# /model arona-grok45

4. Verify

grok -p "Introduce yourself in one sentence" -m arona-grok45

curl https://sub2api.lingshichat.top/v1/responses \
  -H "Authorization: Bearer $ARONA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "input": "ping"
  }'

web_search check

Ask in Grok Build:

Use web_search to find 3 public news headlines from today.

Success: search summaries returned
402 / spending-limit: check web_search model section and supports_backend_search

Field reference

Field	Meaning
`model`	Upstream model ID (`grok-4.5`)
`base_url`	OpenAI-compatible root; include `/v1`
`env_key`	Read key from env instead of plaintext config
`api_backend`	`responses` / `chat_completions` / `messages`
`supports_backend_search`	Required for server-side web_search
`context_window`	Client compaction estimate. Prefer `200000` (quality cliff); hard max ~`500000`, but quality drops past ~200k

Context window (important)

Tier	Tokens	Guidance
Recommended working set	≤ ~200k	Keep coding agents here for stable quality
Hard max	~500k	Accepted by the model, but quality degrades past ~200k
Client `context_window`	Prefer `200000`	Triggers compaction before the quality cliff

Practice:

Set context_window = 200000 — do not default to 500000 just because the spec allows it.
Start a new session / compact mid-task instead of stacking endless history.
Avoid dumping entire monorepos; scope files to the task.
Only raise to 500000 temporarily when you truly need the hard window and accept lower quality.

Optional: Chat Completions backend

api_backend = "chat_completions"

Fine for basic chat; prefer responses + supports_backend_search for web_search.

FAQ

Model list missing arona-grok45

Validate TOML syntax
Restart terminal / Grok Build
Run grok models

401 Unauthorized

Env var visible in the current shell
Full key copied
Key not written to the wrong field

Chat works, web_search fails

[models].web_search points at your relay model
That section’s base_url is the gateway
supports_backend_search = true
Both chat and search use a working grok-4.5

Context compresses too soon / quality drops on long sessions

Prefer quality: set context_window to 200000 so the client compacts near ~200k instead of waiting for 500k.
Grok 4.5 hard max is ~500k, but quality drops past ~200k. Larger context_window is not “smarter”.
If you already set 500000 and the session is huge: new session, drop unrelated history, shrink attached files.
Only use 500000 temporarily for rare ultra-long dumps, then set it back to 200000.

Confirm data training is off

/privacy → training / retention off
/settings Privacy / Data controls
Optional [features] telemetry = false
Account-side Grok training toggles on X if logged in

Do not paste secrets while training options are still on.

Get started

Clients (Grok 4.5)

Guides

API reference

Grok Build setup

Grok Build setup (Grok 4.5)

Recommended settings

0. Disable data training (required)

Optional: disable client telemetry

1. Prepare an API key

2. Edit `~/.grok/config.toml`

Recommended full config

Option B: override built-in `grok-4.5`

Option C: shared endpoint base

3. Launch and switch models

4. Verify

web_search check

Field reference

Context window (important)

Optional: Chat Completions backend

FAQ

Model list missing arona-grok45

401 Unauthorized

Chat works, web_search fails

Context compresses too soon / quality drops on long sessions

Confirm data training is off

​Grok Build setup (Grok 4.5)

​Recommended settings

​0. Disable data training (required)

​Optional: disable client telemetry

​1. Prepare an API key

​2. Edit ~/.grok/config.toml

​Recommended full config

​Option B: override built-in grok-4.5

​Option C: shared endpoint base

​3. Launch and switch models

​4. Verify

​web_search check

​Field reference

​Context window (important)

​Optional: Chat Completions backend

​FAQ

​Model list missing arona-grok45

​401 Unauthorized

​Chat works, web_search fails

​Context compresses too soon / quality drops on long sessions

​Confirm data training is off

​Related

Grok Build setup (Grok 4.5)

Recommended settings

0. Disable data training (required)

Optional: disable client telemetry

1. Prepare an API key

2. Edit `~/.grok/config.toml`

Recommended full config

Option B: override built-in `grok-4.5`

Option C: shared endpoint base

3. Launch and switch models

4. Verify

web_search check

Field reference

Context window (important)

Optional: Chat Completions backend

FAQ

Model list missing arona-grok45

401 Unauthorized

Chat works, web_search fails

Context compresses too soon / quality drops on long sessions

Confirm data training is off

Related